APPING AREA
ed and grouped,
for each group.
transformation a
d. This can be
CP | ] or the
ethods work on
points, therefore
n the ground of
spin-images. In
m is presented.
^a between view
> estimate of the
idual distances
IP (5)
a generic group
1 matrix R and
ernion represen-
this approach, a
malizing by the
ce between two
in be obtained.
ormation would
verage distance,
groups with a
' distance lower
efore one could
g a low number
stimate.
method.
vs with Horn
plemented. For
zh the transfor-
Next, for each
determined. If
esolution, then
common area.
pping zone for
ansformation is
By comparing euclidean distances, above procedure allows to
select the best rototranslation among the whole set of transfor-
mations, obtained by surface matching based on the use of spin-
images. Given only one transformation, a further refinement of
view pair registration can be undertaken, using a global aligne-
ment method , such the ICP algorithm. Since ICP belongs to the
class of greedy algorithms, it needs an initial good approximate
of rotoranslation parameters in order to avoid the convergence
toward a local minimum, rather than to global one. Therefore, in
presented work, another global alignement method was applied,
before the refinement with the ICP. The Frequency Domain
algorithm can be profitable used just to provide such good
initial estimate. Even this algorithm works on common areas
rather than on point correspondences, therefore the overlapping
area between two views must be determined. To that end,
original range data, which make up the two polygonal meshes,
are considered at this step. A procedure like the one previously
described for overlapping area detection, is applied. In this case
the point clouds represented by the range data of the two views
are registered each other, using the previously estimated
transformation. Although it is not very accurate, one can
expects that the applied rototranslation would put points, on the
common area between the two views, very close each other.
Again, this zone can be identified as the one composed by the
set of corresponding points which euclidean distance is less than
a certain threshold. A point on view A, which don't belongs to
the overlapping area, will have its closer corresponding point on
view B at a distance greater than choosed threshold. Also in this
case, the threshold value was set to 1.5 times the mesh
resolution.
In figure 7a-7b an example of common areas detected with
described algorithm is displayed. The point cloud of fig. 7b
represents the data source used to create the mesh of figure 5.
The meshes obtained by both figures were aligned with the
rototranslation parameters estimated by Horn method.
+4 4 8.8.88
EH
8
Li
Fas s
2
EE A
T
" A i
-90 -80 -100 -120
Figure 7a-7b: example of common area detection btw two views
© 2 ... 5 E
s toe : :
**99459mw06e* ° - -
a... : ^
80 fe p putat Doo : s
LE. me yr cere te 6 111130000001 S
k**vb. D$ .. . -
9 Ay v es. =; -
1 se dois :
. &* : -
tpl at $Y. : :
DS esed 9 >
> uS state Sj ME :
E... 7 S 2*?o.$ "» .9. o o =" .....>..
60 * MAPA SIs ;
+ - 283 °°. ee :
fetes 000°...
Ba9,9,9 o 0" * * "^ Doc, 2
PRR EE faite e.
CC ... *. :
Seo ogo 8 Neo = 3 + cy : .
0: --- TRIES 6 Nt Se on oF Ta pe,
0m 00034405 Cadet ses
ES Swen pe. Wert? co?
pass = 30e dire ee
es oies = sges Poe... .
em ak m ve ET EEE
20H ----- bod DEL go CF I) 202 20% cee. 0 *
wt Chr ES “ils 6 ©
T. ce ee SO ... ie
e^ Sevres age +»...
ci. A7 Smet de 0 0
hes tives seo eles ..
ea ga AT SA 00
TO iwi 24330033 :
e $'9*e ap 4$.09-299 =
* ».6 9 e" ese
EL, .. > .
Pe $? $9.2,» 9 £^ e .
AY) TS 20ma 4e 0 8 uma a, «À -
roues t 0,0 ‘2 % Bruce, ew -
—20 rss pose. eei. Loo. a fe. A. $e, 2, 9 P $y
[38 eee - = re y py Neun
edis : 5 — esso P imo P
es ; CET
be fae o ates
Ps 8a) e
FEA ds oo > : e >
—40 375° ze. ne b e eie eU s nis T i. =
> ... «nit 2
sam e se Li ws
? 94 West *
we A m ont rt 8%
oa’ © o° $99, es.
e s)" > Mes.
—60 }- - -- +0 —- so shed vy Met: 0e
9 >.
tee are mee
! e 9 9 S int :
*» e. "uet T qe
eve = ® .
$49 79 5.7
p.c esa
80 ----- "beet uU a
+. *$ ee $4
ex vs e
ae Ses 9099
e c T, a
"aseo oo, i i
20 0 -20 —40 —60
Figure 8: alignement of two views of figure 7, based on
detected common areas.
CONCLUSIONS
Surface matching is the process that compares surfaces in order
to detect similarity among them. It plays a prominent role in 3D
modeling of real objects, as it is used for registration of view
pairs. In this work an alternative procedure for automatic
alignement of view pairs has been presented. The method is
based on the use of an innovative solution in the research field
of automatic registration: spin-images. Spin-images provide a
new kind of object shape encoding, where the global properties
of the object are retained despite of the specific position of
surface points. Since spin-images are constructed with respect
to specific surface points, they are coordinate system indepen-
dent and more discriminant respect with its base points.
Furthermore, as they are constructed without surface fitting,
spin-images can be easily implemented. Given these features,
spin-images represent a tool that can be profitable employed in
registration algorithms.
Thus, on the ground of spin-image concept, a registration
method has been developed. Firstly, point correspondences are
determined, ranking the similarity measure between spin-
images of the poligonal meshes of view pairs. Next the common
area between these views is estimated through the result of
previous step. The method was tested on several real objects,
such statues. Although these objects were characterized by very
complex and irregular shapes, the proposed registration system
was succesfull in all test. These result could be achieved
adopting specific processing strategies, such the analisys of
similarity measure histogram and point correspondences
—319—
